---
title: "The Shortage Was a Financing Decision"
description: "For three years, the defining fact of the AI boom was that you could not buy a Nvidia GPU. The Wall Street Journal now reports the shortage was, at least in part, bought — not found."
date: 2026-11-14
author: Burak
url: https://deepstack.bsormagec.com/p/the-shortage-was-a-financing-decision/
image: https://deepstack.bsormagec.com/images/the-shortage-was-a-financing-decision.webp
---

# The Shortage Was a Financing Decision

*For three years, the defining fact of the AI boom was that you could not buy a Nvidia GPU. The Wall Street Journal now reports the shortage was, at least in part, bought — not found.*

Nov 14, 2026 · Deep Stack by Burak

On Thursday, the Journal published an investigation into how Nvidia handled the other side of the shortage: not the supply of chips, but the demand for them. According to the report, former CEO Jensen Huang used Nvidia's own balance sheet to finance AI startups and smaller cloud providers so they could afford to keep buying Nvidia GPUs — rather than letting an oversupply of chips reach the market and force prices down. The practice has a name in the coverage: circular financing. Nvidia has long rejected the label.

The interesting thing about this story is not that a chipmaker financed its customers. Vendor financing is old. The interesting thing is what the financing was for. It was not for building a market. It was for absorbing a surplus that Nvidia itself had already committed to buying.

## The avalanche

The report centers on a tense meeting in August 2026 between Huang and Nvidia's longtime chief financial officer, Colette Kress — now the interim CEO after Huang resigned on November 12. Kress wanted to rein in the circular-financing commitments to smaller cloud providers and major AI firms such as OpenAI and Anthropic. Huang refused, and reportedly belittled her for questioning the deals, calling her "a deer in the headlights."

Huang's own justification, as the Journal reports it, was blunt. The big cloud providers were reducing their reliance on Nvidia chips, pivoting toward in-house custom silicon. Meanwhile the vast volume of chip wafers and memory Nvidia had already committed to buying from TSMC was, in Huang's words, "coming at him like an avalanche." The company had prepaid for silicon it could not sell into a market that was quietly building alternatives.

So instead of letting the surplus flood the market, Nvidia funded the buyers. By late August, according to the report, the company had paused some of the financing arrangements and introduced additional reviews involving senior executives and board members. It had also contacted TSMC to wind down upcoming chip orders. The avalanche was real enough to renegotiate with the foundry. It was just never allowed to reach the price tag.

## The numbers on the table

The largest proposal in the report involves OpenAI. Early discussions floated Nvidia underwriting roughly 8 gigawatts of OpenAI computing capacity, valued at up to $250 billion. Kress succeeded in cutting the signed guarantee cap to about 4.25 gigawatts, or up to $125 billion. A related Ohio data center project had OpenAI planning to deploy approximately $300 billion worth of Nvidia chips; a revised agreement was signed on August 17.

Those figures sit on top of a balance sheet that had already swollen. SemiAnalysis, reading Nvidia's latest quarterly filing, put the company's total commitments and guarantees at about $530 billion last quarter — up from $184 billion the quarter before. That is a near-tripling in three months, and it is the number that makes the circular-financing charge land. A chipmaker does not accumulate half a trillion dollars in commitments because demand is organic. It accumulates them because demand is being underwritten.

Some of the deals were handled by Nico Caprez, Huang's son-in-law, who was promoted to Nvidia vice president earlier this year, according to VideoCardz's account of the report. The structure was straightforward: help smaller cloud providers and data center companies secure financing to buy Nvidia hardware, and take a share of their revenue in return. Nvidia could help customers afford its GPUs while also benefiting from the sales those deals generated. Everyone in the loop got paid. The people outside the loop paid the price the loop protected.

## The idle warehouse

The other number that matters comes from industry analyst Ed Zitron, who estimates that at least $200 billion worth of Nvidia AI GPUs are sitting idle in warehouses — bought by cloud giants that lacked either the electrical grid capacity or the immediate end-user demand to switch them on. That is an estimate, not an audit, and it should be read as one. But it fits the shape of the Journal's account: chips were bought, the buying was financed, and the machines were not necessarily turned on.

That sounds like a demand problem. It isn't. It is a price problem wearing a demand costume. If the chips had reached the open market, prices would have fallen. Instead the chips reached warehouses, and the prices stayed where the shortage narrative put them. For three years, builders paid scarcity prices for hardware that was, in part, being stockpiled behind a wall of guarantees.

The Hacker News discussion, which ran to more than 570 comments by Thursday afternoon, was notably unimpressed. The dominant tone was skepticism — "we knew a year ago" was the recurring answer, with readers pointing to a Guardian piece from December 2025 that had already raised the circular-financing question, and to the shortseller angle hanging over the whole story. That skepticism is fair, and it is also the point. The market knew the shape of this a year ago. What the Journal added is the inside of it: the meeting, the numbers, the quote.

## The catch

None of this is established fact. That caveat belongs in the body of the article, not in a footnote. The Wall Street Journal's original piece is paywalled and was not read directly for this post; the account here rests on the Wccftech and VideoCardz summaries of it. No Nvidia response appears in the articles read. No funded startup is named. The funding amounts beyond the OpenAI figures are not disclosed. The $530 billion commitments figure comes from SemiAnalysis's reading of a filing, and the $200 billion idle-GPU figure is one analyst's estimate. Treat the story as an allegation with a paper trail, not as a verdict.

There is a second caveat, and it cuts the other way. Even if every detail of the report holds, the financing did not invent the AI boom's demand. Data centers were being built, models were being trained, and the chips were real. Circular financing does not mean the demand was fake. It means the marginal demand — the demand at the edge that kept the shortage narrative alive and the prices high — was financed by the seller. That is a narrower claim than "the shortage was fake," and it is a more useful one. A fake shortage is a scandal. A financed margin is a business model, and business models repeat.

## For builders, the takeaway is practical

The GPU you could not buy, and the rental rate you could not negotiate, were priced against a shortage that Nvidia's own balance sheet was quietly managing. The lesson is not that hardware will suddenly get cheap — Nvidia has already moved to wind down TSMC orders, and the first earnings under Kress, due Tuesday, November 17, will be the next data point. The lesson is that "shortage" is a claim about a market, and markets have balance sheets behind them.

When a vendor tells you the product is scarce, ask who is financing the buyers. When the answer is the vendor, the scarcity is not a fact of nature. It is a financing decision. And financing decisions can be unwound a lot faster than a fab can be built.

That is the problem with a meter you do not own: someone else decides what the reading means.

Source: [Instead Of Lowering GPU Prices, Former NVIDIA CEO Jensen Huang Reportedly Financed AI Startups To Absorb Excess Silicon, Shattering The Shortage Perception](https://wccftech.com/jensen-huang-financed-ai-startups-absorb-excess-silicon/amp/)